Recipe

Planner + executor agent split

Separate reasoning from action. The planner decides what to do; the executor carries it out with tool access and a tight feedback loop.

Overview

This recipe splits a single agent into two cooperating roles. The planner receives the user request, breaks it into a sequence of discrete steps, and hands the plan to the executor. The executor runs each step, calling tools as needed, and reports results back. If a step fails, the planner revises the remaining plan.

When to use

Multi-step tasks where tool selection depends on prior results.
Expensive reasoning models — keep the planner small and focused.
Tasks that benefit from a human-readable plan before execution.
Long-running workflows where the plan may need mid-flight revision.

Structure

User → Planner
         ↓ plan
       Executor
         ↓ tool calls
       Tools (browser, code, search)
         ↓ result
       Executor → Planner (revise?)
         ↓ final
       User

Key parameters

Parameter	Default	Notes
max_plan_steps	8	Hard cap on plan length
executor_retries	2	Retries per step before replan
planner_model	sonnet	Model used for reasoning

Next steps

Pair this recipe with the tool router pattern for dynamic tool selection, or add a human-in-the-loop gate before the executor runs expensive operations.